8 months without a post, the AI bubble debate, and how my engineering job actually changed

How AI contributed to this article (click to expand)
  • I recorded a raw voice memo detailing my observations over the past eight months.
  • I used Hermes Agent to transcribe and structure my points into markdown.
  • All opinions, personal experiences, and engineering conclusions are strictly my own.

Rationale for creating this section explained here


The bubble debate: scams vs $30T forecasts

I haven't written on this blog in eight months. I’m a terrible blogger when it comes to consistency, but lately the public debate around AI has become so unhinged that I wanted to get a few things off my chest.

Ed Zitron spends half his time on Where's Your Head At insisting the entire generative AI industry is a fraud built on sand. Meanwhile, Anthropic and Wall Street bankers pitch over $30 trillion in addressable market potential to prospective investors (recently reported by the Wall Street Journal and analyzed in Fortune).

Both takes are ridiculous. I am a software and data engineer building and maintaining production data pipelines every single day. Calling the technology a scam ignores what is actually happening in our repositories, and pretending it is an immediate $30 trillion economy is pure marketing theater. Here is my honest take on what I am seeing on the ground.

Google's TPU vertical integration

Until late last year, I viewed coding models mostly as glorified autocomplete for boilerplate. November changed my perspective completely.

Google's release of Gemini 3 Pro was the first model that genuinely surprised me with its engineering accuracy. The real story behind it was the hardware. Google trained and served Gemini 3 entirely on their custom TPU infrastructure, leaving Nvidia GPUs completely out of the equation.

Everyone spent two years crowning Nvidia as the undisputed king of AI. Google proved that vertical integration still works, and they are currently the only player with custom silicon, datacenter power, research talent, and distribution under one roof.

Anthropic and enterprise adoption

Anthropic followed up with Claude Opus 4.5 in December, and by February, Claude was dominating engineering tasks on both the LMSYS Chatbot Arena and Artificial Analysis.

I saw the impact firsthand at work. Claude became our team's default model practically overnight. With reported quarterly revenues heading toward $40 billion annually, enterprise teams are writing real checks because the model saves real hours.

Writing code vs reviewing code

Skeptics can write thousands of words about why AI is a fad, but here is my daily reality: I have not written code by hand since March.

I spent fifteen years building my craft as a Python and data engineer. Today, my fingers rarely touch the keyboard to type raw syntax.

I spend my hours reading pull request diffs, checking edge cases, and reasoning about system design. An AI companion writes the files, runs local tests, and creates the pull requests.

The argument that AI cannot write software is completely dead. The real unknown is how long human oversight remains a bottleneck in the loop.

Model harnessing and autonomous agents

Raw models in a chat window are clumsy. The real leap this year came from model harnessing.

Tools like OpenCode showed how much faster work gets when a model can run terminal commands directly. Agent frameworks like Hermes Agent (from Nous Research) took that concept to its logical conclusion.

I use Hermes every single day on my personal projects. It reads my repos, triages my emails via Gmail APIs, conducts background research, and handles Git operations directly. Having an agent that actually acts on your environment changes everything.

Open weights vs closed moats

Back in 2023, Jeremy Howard wrote AI Safety and the Age of Disenlightenment, arguing that open weights are essential for society. He was completely right.

Look at what happened with DeepSeek and Alibaba's Qwen. Open-weight models from China are now matching closed Western models on coding benchmarks.

Anthropic’s narrative that only a tiny club of "responsible" labs should hold frontier models feels like self-serving corporate lobbying. Trying to rebrand closed moats as "safety" is alienating developers.

Nvidia gets this. Their deep backing of Hugging Face proves that hardware sales depend on a thriving open ecosystem where developers can download and run models anywhere.

Google Flash models and latency economics

While people were waiting for mythical giant models, Google did something smarter with their Flash line (now on 3.8).

Checking the Artificial Analysis Speed & Pricing Index, Gemini Flash is roughly 10x faster and an order of magnitude cheaper than flagship models, while being plenty smart for web dev and ETL pipelines. When you build autonomous agent loops, latency and token costs matter ten times more than winning a theoretical benchmark by 1%.

AI capex and upcoming IPOs

Pitching a $30 trillion market is fantasy math. Even at today's real numbers, the circular economics are wild.

If Anthropic pulls in $40 billion in revenue with zero profit, almost every dollar flows right back into Google Cloud, TPUs, and data center capex.

Nobody knows if these unit economics actually work in the long run. We won't have real answers until OpenAI or Anthropic goes through an IPO and faces public earnings calls. Until then, expect plenty of creative accounting between hyperscalers and labs to make margins look respectable.

Automated vulnerability discovery

One area that deserves way more attention is automated vulnerability discovery.

Between internal safety red-teaming and research from groups like Google Project Zero, LLMs are becoming ruthlessly effective at fuzzing open-source codebases and finding memory flaws.

The traditional market for selling zero-day exploits is in serious trouble. When automated models can continuously find and patch security bugs before humans can weaponize them, offensive security economics collapse.

Databricks Genie in production

In my day job, our data stack runs on Databricks, and using Databricks Genie has completely reshaped our workflow.

We pair it with GitHub and the Model Context Protocol (MCP):

  1. We describe the business logic and goals in natural language.
  2. The agent inspects schemas, generates ETL transformations, and drafts unit tests.
  3. The pull request opens automatically.
  4. We review, verify the edge cases, and merge.

You can argue about bubbles and valuations on Twitter all day. Inside real engineering teams, the transition happened months ago. We aren't going back.